YandexGPT Integration in GraphRAG: Graph Construction and Semantic Analysis of Graph Communities

Knowledge graphs are powerful tools for organizing and extracting information from unstructured text. While GraphRAG – a method combining retrieval-augmented generation with graph-based reasoning – has achieved increasing success in tasks like question answering and recommender systems in English-speaking domains, its implementation with Russian-language data remains relatively underexplored. This study focuses on adapting the GraphRAG framework for Russian by integrating it with YandexGPT, one of the most advanced large language models trained on Russian corpora. A key contribution of this work is the re-engineering of the original GraphRAG codebase to support API-based access to proprietary models such as YandexGPT. Additionally, we introduce an interactive web interface that enables global semantic search across the constructed KGs. To benchmark performance, we also conduct comparative experiments on Russian datasets of podcast transcriptions and news articles with open-source models, such as Qwen, DeepSeek et al., which are primarily trained on English data. We expect that YandexGPT, being more finely tuned to the Russian language, will produce graphs that are both more precise and more informative. Our evaluation pipeline combines established and novel metrics, integrating semantic aspects – comprehensiveness, directness, diversity, and empowerment – with structural measures like degree centrality. We enhance this analysis by using external LLMs with extended context windows to provide a more holistic assessment of graph completeness and relevance. Keywords: GraphRAG, YandexGPT, knowledge graphs, Russian datasets, semantic relationships, NLP, LLM